Is Granger causality analysis a linear operation？

[Francois]

@NingZhuang

(5) For each subject i, take the average of Cij under all trials, Di = average (Cij) = average (Aij-Bij). Di is the difference matrix corresponding to the final two conditions obtained by each participant.
(6) Di was tested by "Process1-Test-Parametric Test against zero "for all subjects.

The recommended approach would be to use non-parametric tests A vs B.
https://neuroimage.usc.edu/brainstorm/Tutorials/Statistics#Nonparametric_permutation_tests

I found that if they were far apart in space, the values in their connection matrix would be very small.

You would observe this as well with a simple correlation. The minimum norm model produces very smooth distributions of activities, the source signals are linear recombinations of all the sensors signals, ans therefore all the source signals are correlated with each other. The closer they are, the more correlated they are.

I would recommend you do not spend too much time trying to interpret the Nscouts x Nscouts matrices of GC values, or trying to compare the GC between different pairs or ROIs. The interesting result is the output of the non-parametric test: what is significantly different in these connectivity matrices between your 2 conditions.